The technology of analysis in the figurative problems of speech processing digital information. Scientific Visualization Volume 5, Issue 3, 2013, Pages 75-88

This work is devoted to the research of the pattern analysis-synthesis possibilities for speech signals spectrograms images in the different areas of application: speech encoding, noise or distortion canceling, speaker identification, speech compression and etc. The different algorithms of sound signals synthesis on the predetermined spectrogram image basis are described. The comparative quality analysis for different synthesis algorithms on the basis of the: whole sonogram, local maximums, divisible to the main tone harmonics with the natural and synthesized phase are presented. The quality analysis was carry out taking into account the following algorithms characteristics: the algorithm performance and the difference rate between the original and the synthesized sonograms. With the aim to realize the difference rate quantitative measurement the notion "difference norm" is introduced. All mentioned above synthesis algorithms have been embodied in a single software package "SoundTool", which was developed using the parallel programming technology Nvidia CUDA in order of performance improvement. In addition, this software package also allows: the sound signal sonogram editing, sonogramm image import and export into standard graphical editors, median filtering for narrowband noise canceling, in particular, the electric power supply noise canceling.
Indexed keywords
Algorithm performance; Fourier transformations; Noises; Programming technology; Quantitative measurement; Sonogram; Sound visualization; Speaker identification